NWTOPT – A hyperparameter optimization approach for selection of environmental model solver settings. (January 2022)

• Record Type: Journal Article
• Authors: Newcomer, Max W.; Hunt, Randall J.
• Is Part Of: Environmental modelling & software. Volume 147(2022)
• Publication Date: 2022-01
• Subjects: Groundwater modeling -- Hyperparameter -- Optimization -- Parameter estimation -- Uncertainty -- Calibration; Environmental monitoring -- Computer programs -- Periodicals; Ecology -- Computer simulation -- Periodicals; Digital computer simulation -- Periodicals; Computer software -- Periodicals; Environmental Monitoring -- Periodicals; Computer Simulation -- Periodicals; Environnement -- Surveillance -- Logiciels -- Périodiques; Écologie -- Simulation, Méthodes de -- Périodiques; Simulation par ordinateur -- Périodiques; Logiciels -- Périodiques; Computer software; Digital computer simulation; Ecology -- Computer simulation; Environmental monitoring -- Computer programs; Periodicals; Electronic journals; 363.70015118
• DOI: 10.1016/j.envsoft.2021.105250
• Languages: English
• ISSNs: 1364-8152

Abstract

Abstract: Hyperparameter optimization approaches were applied to improve performance and accuracy of groundwater flow models. Freely available new software, NWTOPT, is described that uses Tree of Parzen Estimators (TPE) and Random Search algorithms to optimize MODFLOW-NWT's solver settings. We ran 3500 trials on a steady-state and transient model. To quantify the performance of candidate solver settings, we defined a loss function based on time elapsed and mass balance error of the MODFLOW-NWT forward run. Before optimization the steady-state model ran in ∼12 min and the transient model ran in ∼5 h with acceptable mass balance error (<1%). After optimization runtimes were reduced to ∼2.7 min (steady state) and ∼48 min (transient) with errors below 0.1%. In both cases TPE found hyperparameters that resulted in faster running and lower error models than those found by Random Search. The time to complete the optimization trials was also shorter with the TPE algorithm. Highlights: NWTOPT customizes off-the-shelf software for seamless application to MODFLOWNWT solver settings. NWTOPT uses high-throughput computing to test multiple solver settings simultaneously. The NWTOPT performance metric "loss" captures both aspects of desirable MODFLOW-NWT runs – short run times and low mass balance errors. Benefits over manually selected solver settings are demonstrated using both a steady state and transient MODFLOW-NWT model. The Tree of Parzen Estimator algorithm performed better than a Random Search approach for both steady state and transient models.